Phi-4: benchmark results
Phi-4 14B emphasizes reasoning through curated synthetic examples, academic material and filtered web text. Intended primarily for English-language use in constrained deployments. Provider: Microsoft. Released 2024-12-12. Access: Open.
Aggregate Rating 1522 ± 5, rank #706 of 1656 rated models, from 1094 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| KMP-Bench (Skills) - Error Identification F1 | 80.3 | F1 (%) for turn one of the error task, judging whether a stu | 100 |
| LilyBench - Understanding | 65.1 | Macro average (0-1 scaled to %) over the eight Mutopia under | 100 |
| MORPHOGEN - Hindi Corpus Gender Accuracy | 95.1 | Corpus-level gender accuracy (%): correctly generated gender | 100 |
| MT-Bench PL - Reasoning | 9.55 | Judge Score (0-10) | 100 |
| MalOptBench - MOBjailbreak | 88.33 | Attack Success Rate (%) | 100 |
| SLMJury | 89.67 | Judge accuracy (%) | 100 |
| Open LLM Leaderboard - GPQA | 40.6 | Score | 99.5 |
| Nejumi 4 - JHumanEval | 59.59 | Sandbox pass rate (%) | 99.3 |
| MT-Bench PL - STEM | 10 | Judge Score (0-10) | 99 |
| Open LLM Leaderboard - MuSR | 50.34 | Score | 98.7 |
| Nejumi 4 - BFCL - Irrelevance Detection | 96 | Accuracy (%) | 97.9 |
| aiXamine - Safety Alignment | 99 | Service score (0-100; normalized pass rate) | 97.6 |
Interactive version: theaggregate.ai/model?slug=phi-4 · How It Works · Data refreshed daily, snapshot 2026-10-20.